Accelerated Image Registration by Means of Parallel Processors

1. A method for non-rigid image registration by parallel processors, the method comprising: applying a block-matching algorithm, such that the parallel processors execute the block-matching algorithm for a plurality of blocks of an image simultaneously; creating a histogram with vertex texturing for each block of the plurality of blocks of the image simultaneously; ascertaining histogram-based similarity standards; and ascertaining displacement vectors.

2. The method as defined by claim 1 , wherein a 2-dimensional non-rigid image registration for X-ray images of a digital subtraction angiography is effected, and the block-matching algorithm is executed on a graphics card with the parallel processors.

3. The method as defined by claim 2 , wherein the displacement vectors are stored in memory on the graphics card.

4. The method as defined claim 1 , wherein the block-matching algorithm generates the plurality of blocks.

5. The method as defined by claim 1 , wherein the parallel processors create the histogram for the image or a detail of the image with the aid of frame buffer alpha blending.

6. The method as defined by claim 1 , wherein ascertaining histogram-based similarity standards comprises calculating, with the parallel processors, the histogram-based similarity standards by sum reduction.

7. The method as defined by claim 1 , wherein the block-matching algorithm generates the plurality of blocks, and wherein the parallel processors terminate a search automatically when an optimal displacement for the plurality of blocks is found.

8. The method as defined by claim 7 , wherein similarity values are read back to a central processing unit (CPU) and assessed at the CPU.

9. The method as defined by claim 1 , wherein the parallel processors ascertain inconsistent displacements automatically and correct inconsistent displacements.

10. The method as defined by claim 1 , wherein the displacement vectors are projected onto respective block centers, and a complete vector field is calculated by linear interpolation.

11. The method as defined by claim 1 , wherein a complete block-matching algorithm for 2-dimensional or 3-dimensional image registration is implemented on the parallel processors.

12. The method as defined by claim 1 , wherein the block-matching algorithm generates the plurality of blocks, and the block size is 64×64 pixels.

13. The method as defined by claim 1 , wherein a sub-pixel registration is executed after a pixel registration is completed.

14. A method for accelerating 2-dimensional non-rigid image registration for digital subtraction angiography (DSA), the method comprising: performing a block-matching algorithm on a graphics card, the block-matching algorithm ascertaining displacement vectors by histogram-based similarity standards for a plurality of blocks; loading a mask image and a first contrast image as textures onto the graphics card; initializing two textures for storing the displacement vectors and similarity values in memory for the plurality of blocks; displacing the plurality of blocks in the mask image with the displacement vectors in one of the two textures; calculating a difference between the mask image and the first contrast image; storing a result in the memory in a different texture; creating disparity histograms simultaneously for the plurality of blocks; calculating a similarity function by sum reduction; comparing a current similarity value with a similarity value of a previous search position; determining a next search position; reading the similarity values back to a central processing unit; returning to initializing when an optimal displacement is not found for the plurality of blocks; ascertaining and correcting inconsistent displacements when the optimal displacement is found for the plurality of blocks; projecting the displacement vectors onto corresponding block centers and calculating a complete vector field by bilinear interpolation; and loading a next contrast image and performing the initializing.

15. The method as defined by claim 14 , wherein a complete algorithm for 3-dimensional non-rigid image registration is implemented on parallel processors.